Adaptability evaluation method for virtual reality building design

Abstract

The invention relates to the technical field of building design, and discloses a virtual reality building design-oriented adaptability evaluation method, which comprises the steps of collecting building design data through a virtual reality environment; dynamically updating the adaptability evaluation model by utilizing an incremental learning framework, wherein the dynamic updating comprises knowledge retention based on historical tasks and incremental learning of new tasks; the generative adversarial network is adopted to expand the original building design data, and the diversity and quality of the generated data are improved through a composite loss function; training a neural network based on a self-adaptive force field optimization mechanism, and dynamically adjusting parameter updating strength; and carrying out adaptability evaluation on the trained model, and outputting comfort levels of the design scheme in different scenes. According to the method, the generative adversarial network of the condition generator, the multi-task discriminator and the composite loss function is combined, the building design data is expanded, and the problems of insufficient original data samples and unbalanced categories are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of architectural design, and particularly to an adaptability evaluation method for virtual reality architectural design. Background Art

[0002] The adaptability evaluation of architectural design is a key link in the architectural design process. Its goal is to evaluate the adaptability and performance of a design scheme in different usage scenarios by analyzing multi-dimensional features such as the geometric form, material properties, and layout configuration of the architectural design scheme. However, traditional adaptability evaluation models have certain limitations when facing dynamically updated architectural design data.

[0003] The main manifestations are as follows: Architectural design data is usually collected in real time, with strong data dynamics. The introduction of new data may cause the model to forget the knowledge of previous tasks; the amount of collected architectural design data is limited, and the data distribution of different categories is unbalanced, which is likely to cause the problem of insufficient samples in model training; in addition, traditional model training methods are prone to falling into local optimal solutions when dealing with high-dimensional complex features, and it is difficult to fully extract the complex characteristics in architectural design.

[0004] With the wide application of virtual reality technology in architectural design, it has become possible to collect architectural design data through user behavior and interaction, providing more diverse evaluation dimensions, such as user behavior paths, gaze tracking, etc. However, these data features are complex and diverse, posing higher requirements for the dynamic adaptability and generalization ability of the model. At the same time, the evaluation model needs to have the ability of continuous learning to cope with the dynamic changes of the design scheme and avoid the high computational cost brought by retraining the entire model due to data updates. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention proposes an adaptability evaluation method for virtual reality architectural design to solve the above technical problems.

[0006] The adaptability evaluation method for virtual reality architectural design includes:

[0007] Collecting architectural design data through a virtual reality environment, where the architectural design data includes user interaction behavior paths, geometric form parameters, material property parameters, and environmental simulation parameters;

[0008] Dynamically updating the adaptability evaluation model using an incremental learning framework, including knowledge retention based on historical tasks and incremental learning of new tasks;

[0009] Adopting a generative adversarial network to expand the original architectural design data, and improving the diversity and quality of the generated data through a composite loss function;

[0010] Train a neural network based on an adaptive force field optimization mechanism to dynamically adjust the intensity of parameter updates;

[0011] Conduct an adaptability assessment on the trained model and output the comfort level of the design solution under different scenarios.

[0012] Furthermore, the acquisition methods of the building design data include automated export and semi-automated interactive recording; the automated export extracts data from the design software automatically through a design software tool; the semi-automated interactive recording realizes the recording and sorting of data through the combination of user interaction and data processing algorithms.

[0013] Furthermore, the specific steps of the incremental learning framework include:

[0014] Judge whether the accuracy rate of the new data set on the validation set of the original data has increased by more than a preset threshold. If it is satisfied, the new and old data are weighted and fused by category;

[0015] Label the unlabeled new data through the pseudo-labeling method to generate an extended training set;

[0016] Iteratively update the model parameters until convergence.

[0017] Furthermore, the input of the conditional generator of the generative adversarial network includes random noise and the feature statistical information of the building design data.

[0018] Furthermore, the composite loss function includes adversarial loss, content loss, and feature loss,

[0019] Among them, the content loss constrains the pixel space consistency between the generated data and the real data, and the feature loss constrains the consistency of the high-level feature space.

[0020] Furthermore, the adaptive force field optimization mechanism adjusts the update amplitude of each parameter through dynamic force field distribution, and the force field distribution adaptively adjusts the interference coefficient and the incremental adjustment factor based on the current parameter state and the change rate of the loss function.

[0021] Furthermore, the adjustment of the force field distribution includes:

[0022] Calculate the initial force field distribution function, combine the non-linear mapping of weights and biases, and use an activation function to realize the non-linear mapping of weights and biases;

[0023] Dynamically adjust the interference coefficient according to the change rate of the loss function to suppress parameter oscillation;

[0024] Adaptive adjustment of the learning rate through the incremental adjustment factor and the perturbation correction factor.

[0025] Furthermore, the output categories of the adaptability evaluation model include low comfort, medium comfort, and high comfort.

[0026] Furthermore, during the training process of the generative adversarial network, the weights of the content loss and the feature loss are gradually increased. In the initial stage, the adversarial loss is the main focus, and in the later stage, the regulatory role of the content and feature losses is gradually enhanced.

[0027] Furthermore, the training termination condition of the adaptive force field optimization mechanism is to reach the preset maximum number of iterations or the change rate of the loss function is lower than the threshold.

[0028] The invention adopting the above technical solution has the following advantages:

[0029] 1. The invention combines a conditional generator, a multi-task discriminator, and a generative adversarial network with a composite damage function to expand building design data and alleviate the problems of insufficient original data samples and class imbalance.

[0030] 2. The invention enhances the similarity between the generated data and the real data in multiple dimensions through content loss and feature damage constraints, thereby improving the quality of building design data.

[0031] 3. The invention adopts an adaptive force field optimization mechanism to dynamically adjust the update intensity of each parameter through the force field distribution, solving the problem of excessive consistency in the parameter update direction and amplitude in traditional neural network optimization, which leads to the local optimum problem.

[0032] 4. The invention combines the incremental adjustment factor and the perturbation correction factor with the dynamic change rate of the loss function to further optimize the learning rate adjustment, enabling the model to better adapt to the complexity of different building design data features. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific embodiments of the present invention, the drawings required for the specific embodiments will be briefly introduced below. In all the drawings, the elements or parts are not necessarily drawn to scale.

[0034] Figure 1 It is a flowchart of the adaptability evaluation method for virtual reality building design of the present invention;

[0035] Figure 2 It is a framework diagram of incremental learning in the adaptability evaluation method for virtual reality building design of the present invention;

[0036] Figure 3 It is a working flowchart of the training method of the adaptability evaluation model in the adaptability evaluation method for virtual reality building design of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The embodiments of the technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, so they are only examples and cannot be used to limit the protection scope of the present invention.

[0038] As Figures 1 to 3 shown, the adaptive evaluation method for virtual reality-based architectural design of the present invention includes:

[0039] Step S01: Collect architectural design data through a virtual reality environment, where the architectural design data includes user interaction behavior paths, geometric form parameters, material property parameters, and environmental simulation parameters;

[0040] Step S02: Dynamically update the adaptive evaluation model using an incremental learning framework, including knowledge retention based on historical tasks and incremental learning of new tasks;

[0041] Step S03: Augment the original architectural design data using a generative adversarial network, and improve the diversity and quality of the generated data through a composite loss function;

[0042] Step S04: Train a neural network based on an adaptive force field optimization mechanism to dynamically adjust the parameter update strength;

[0043] Step S05: Conduct an adaptive evaluation on the trained model and output the comfort level of the design solution in different scenarios.

[0044] In some embodiments, the collection methods of architectural design data include automated export and semi-automated interaction recording; automated export extracts data from the design software automatically through design software tools; semi-automated interaction recording realizes the recording and collation of data through the combination of user interaction and data processing algorithms.

[0045] Specifically, the training architectural design data for training the adaptive evaluation model is collected from architectural design solutions, aiming to evaluate the adaptability of the design solutions through user interaction in a virtual reality environment;

[0046] The collection methods of architectural design data are divided into two types: automated and semi-automated:

[0047] Automated collection involves using software tools to directly export data from the architectural design database and the virtual reality environment system;

[0048] The semi-automated method manually marks and records architectural design data through the behavior and feedback of users in the virtual reality environment.

[0049] The storage format of the architectural design data adopts a structured JSON file, which is convenient for subsequent processing and analysis.

[0050] In a specific embodiment, the attributes of the building design data include:

[0051] R a is the building geometric form, such as volume, height, and shape, etc.; D a is the material property, including material type, durability, and aesthetic degree, etc.; L a is the building layout, detailing the configuration and function division of rooms; I a is the lighting design, such as the configuration of natural light and artificial light and its influence; V a is the visual tracking building design data, recording the stay and movement of the user's line of sight in the virtual environment; P a is the behavior path building design data, tracking the movement trajectory of the user in the virtual environment; O a is the operation record, including the interaction operations of the user on various elements in the virtual environment; W a is the simulated weather condition, such as the simulated building design data of temperature, humidity, wind speed, etc.; E a is the surrounding environment impact, involving the impact of the surrounding buildings and natural environment; S a is the system performance building design data, recording the performance of the VR system during operation.

[0052] The Word2Vec algorithm is used to vectorize the text. According to a preset large-scale corpus, these texts to be vectorized are scanned, and each word is represented as a one-hot encoded vector, and the dimension of the one-hot encoded vector is equal to the size of the vocabulary in the corpus.

[0053] This embodiment is only a data format of the present invention. In practical applications, the attributes of building data are usually more than 10, and the number of attributes of building design data may reach dozens or even hundreds.

[0054] Furthermore, the collected building design data is labeled (manual labeling in this embodiment)

[0055] The categories of labels include:

[0056] 0 - Low comfort level (not suitable for long-term residence or use);

[0057] 1 - Medium comfort level (basically meeting the residence or use requirements);

[0058] 2 - High comfort level (optimizing the residence or use experience).

[0059] In some embodiments, the specific steps of the incremental learning framework include:

[0060] Judge whether the accuracy rate of the new data set on the validation set of the original data has increased by more than a preset threshold. If satisfied, the new and old data are weighted and fused by category;

[0061] Label the unlabeled new data through the pseudo-labeling method to generate an extended training set;

[0062] Iteratively update the model parameters until convergence.

[0063] Specifically, incremental learning can acquire knowledge from old adaptive evaluation model inference tasks, enabling the adaptive evaluation model to learn to solve new inference tasks while retaining the knowledge learned in previous model inference tasks, avoiding retraining the parameters of the adaptive evaluation model when new building design data arrives.

[0064] Based on the training framework of the adaptive evaluation model with incremental learning, input the real-time collected building design data into the adaptive evaluation model, enabling the adaptive evaluation model to learn the features in the real-time collected building design data and, on the basis of maintaining the original capabilities of the adaptive evaluation model, learn new knowledge and capabilities from the real-time building design data. The incremental learning framework is as Figure 2 shown;

[0065] Incremental learning is a continuous learning process. During training, assume that the adaptive evaluation model has learned the first m tasks. When facing a new task (the (m + 1)-th task) T m+1 and other corresponding building design data D m+1 , the adaptive evaluation model trained with historical building design data can use the prior knowledge learned from old model inference tasks to assist in the learning of new model inference tasks, and then update the adaptive evaluation model using the learned knowledge.

[0066] During training, for the newly collected unlabeled building design data, use the pseudo-labeling method to use the unlabeled building design data to assist the original labeled building design data in training. As Figure 3 shown, use the adaptive evaluation model to predict the unlabeled building design data of the new task T i , and add the prediction result R i as a pseudo-label to the original training set to form a new training set D i * , and retrain the network again. In this way, new categories can be gradually learned and the performance of the adaptive evaluation model can be improved.

[0067] The specific workflow of the training method of the adaptive evaluation model based on incremental learning is as follows:

[0068] 1. Import a basic adaptability evaluation model obtained after offline training with the previous original architectural design data. At the same time, determine whether the new architectural design dataset for the next moment is available. The basis for the determination is: whether the accuracy of the adaptability evaluation model in the original architectural design data has improved after being trained with the new architectural design data. If there is an improvement, the new architectural design dataset is available and proceed to the next step; if there is no improvement, the new architectural design dataset is not available and stop the training.

[0069] 2. Select m upd samples for each category of samples in the new architectural design dataset and save them. m upd is a manually preset sample quantity value. Obtain the feature centers by clustering the samples of the same category in the new and old architectural design data. Taking the distance between the two feature centers as the sphere radius at the two feature center points, the new architectural design data points falling inside the two spheres are the selected sample points. Subsequently, add the selected sample points to the same category in the original architectural design dataset, send the saved fused architectural design dataset into the new task, and simultaneously proceed to the next step.

[0070] 3. Start training using the fused architectural design dataset after adding the new architectural design data samples to the adaptability evaluation model to generate the adaptability evaluation model corresponding to the next task. Determine the availability of the architectural design dataset corresponding to the next task. If it is available, return to step 2; if it is not available, proceed to the next step.

[0071] 4. Terminate the training, update the parameters of the adaptability evaluation model, and output the current version of the adaptability evaluation model.

[0072] In some embodiments, the input of the conditional generator of the generative adversarial network includes random noise and the feature statistical information of the architectural design data.

[0073] In some embodiments, the composite loss function includes adversarial loss, content loss, and feature loss.

[0074] Among them, the content loss constrains the pixel space consistency between the generated data and the real data, and the feature loss constrains the consistency in the high-level feature space.

[0075] In some embodiments, the generative adversarial network gradually increases the weights of the content loss and the feature loss during the training process. In the initial stage, the adversarial loss is the main one, and the regulatory roles of the content and feature losses are gradually enhanced in the later stage.

[0076] Specifically, a generative adversarial network is used to achieve diverse building design data augmentation. By adopting conditional information, a multi-task learning discriminator, and a composite loss function in the generator, the similarity between the generated building design data samples and the original building design data in multiple dimensions is enhanced, and the balance of building design data is effectively improved, achieving a more realistic and diverse building design data augmentation effect.

[0077] The training process of the generative adversarial network is as follows:

[0078] 1. Initialize the network parameters of the generator and discriminator of the generative adversarial network, including the weight parameters and bias parameters of the generator and discriminator. The initialization method is random initialization, and the initialized parameters follow a normal distribution with a mean of 0 and a variance of the identity matrix.

[0079] 2. During the process of the generative adversarial network, the generator and discriminator are alternately trained. When training the generator, the generator receives random noise and conditional information as inputs, and the discriminator is mainly used to determine whether the input building design data sample is a real sample or a generated sample. The loss function of the generator is expressed as:

[0080]

[0081] In the formula, L G is the loss function of the generator, denotes expectation, z c is a random noise vector, ∼ indicates following a specific distribution; p z (z c ) is the noise distribution, specifically using a normal distribution; D c () is the discriminator function, which outputs a probability value after receiving the generated building design data sample of the generator, indicating whether the building design data sample is a real building design data sample; is the generated building design data sample of the generator; G () is the generator, which receives the noise z c and the conditional information c c as inputs and outputs the generated building design data sample c ; λ c is the weighting coefficient of the content loss, used to adjust the consistency of the generated building design data sample in the pixel space; c c is the conditional information;

[0082] For example, information such as the feature mean and feature variance of the building design data; X c is the real building design data sample; ∥∥ is the L2 norm.

[0083] Furthermore, the output calculation method of the generator is expressed as:

[0084] G c (z c ,c c ) = f gen (W c z c +b c ,W cond c c )

[0085] In the formula, f gen () is the non-linear activation function of the generator, W c and b c are the weight and bias of the generator with respect to the random noise z c respectively, and W cond is the weight related to the conditional information c c .

[0086] Moreover, when training the generator, the architectural design data samples generated by the generator and the real architectural design data samples are input into the discriminator together to improve the quality of the generated architectural design data, and are constrained by the content loss, and the calculation method is expressed as:

[0087]

[0088] In the formula, L content is the content loss function, i c is the eigenvalue index of the architectural design data. By summing the differences at the i-th pixel position, it is ensured that the generated architectural design data samples and the real architectural design data samples maintain a high degree of consistency in the pixel space.

[0089] Furthermore, in order to ensure the consistency of the high-level feature space, the feature loss is used for constraint, and the quality of the generated architectural design data samples is further improved by constraining the similarity of the high-level features. The calculation method is expressed as:

[0090]

[0091] In the formula, L feature is the feature loss function, is the feature output by the generator network at the k c -th layer, and are the feature representations of the real architectural design data samples and the generated architectural design data samples at the k c -th layer respectively, and k c is the layer index of the generator network.

[0092] 3. When training the discriminator, real building design data samples and the building design data samples generated by the generator are used to improve the overall recognition accuracy and stability of the discriminator. The loss function of the discriminator improves the discrimination accuracy of the discriminator by maximizing the recognition accuracy of real building design data samples and minimizing the misrecognition of the building design data samples generated by the generator. The calculation method is expressed as:

[0093]

[0094] In the formula, L D is the loss function of the discriminator, and p data (X c ) is the distribution of real building design data.

[0095] Furthermore, the calculation method of the output probability of the discriminator is expressed as:

[0096] D c (X c ) = Sig(W d ·Φ feat (X c ) + b d )

[0097] In the formula, Sig() is the Sigmoid activation function, W d and b d are the weights and biases of the discriminator, and Φ feat (X c ) is the feature representation extracted by the discriminator for the input X c .

[0098] 4. Calculate and optimize the loss function. By fusing the adversarial loss, content loss, and feature loss with a composite loss function, the consistency problem of the generated building design data samples at multiple scales is solved, and the quality of the generated building design data samples is improved from multiple dimensions, expressed as:

[0099] L c = L adv + λ c L content + γ c L feature

[0100] In the formula, L c is the total loss function of the generative adversarial network; L adv is the adversarial loss function, L adv = L D + L G ; L content is the content loss function, L feature is the feature loss function, λ c and γc They are the weighted coefficients of content loss and feature loss respectively, which are used to balance different loss terms during training.

[0101] Furthermore, in order to focus on the adversarial loss in the initial stage of training and gradually enhance the role of content loss and feature loss in the later stage, the calculation method of the weighted coefficients of content loss and feature loss is expressed as:

[0102]

[0103]

[0104] In the formula, λ c (t) and γ c (t) are the dynamic weighted coefficients of content loss and feature loss at the t-th iteration respectively. λ min and λ max are the minimum and maximum values of the content loss weight respectively, and γ min and γ max are the minimum and maximum values of the feature loss weight respectively. int(t) is the current training round, and T is the total number of training rounds.

[0105] 5. Use the backpropagation algorithm to alternately update the generator and the discriminator to improve the model stability and the quality of the generated building design data samples. The parameter update method of the generator is expressed as:

[0106]

[0107] In the formula, θ G are the parameters of the generator, including the weight and bias parameters of the generator; ← represents the parameter update operation, η is the learning rate of the generative adversarial network, is the gradient of the total loss function of the generative adversarial network with respect to the generator parameters.

[0108] Furthermore, in order to continuously improve the recognition accuracy of the discriminator for the generated building design data samples of the real generator, the parameter update method of the discriminator is expressed as:

[0109]

[0110] In the formula, θ D are the parameters of the discriminator, including the weight and bias parameters of the discriminator; is the gradient of the total loss function of the generative adversarial network with respect to the discriminator parameters.

[0111] 6. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed.

[0112] In a specific embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0113] After the training of the building design data augmentation model is completed, the trained building design data augmentation model is used to increase the number of building design data samples.

[0114] In a specific embodiment, if the original collected building design data samples are 800, and the building design data augmentation model generates 200 building design data samples through augmentation, then the augmented building design data set contains 1000 building design data samples.

[0115] In some embodiments, the adaptive force field optimization mechanism adjusts the update amplitude of each parameter through dynamic force field distribution. The force field distribution adaptively adjusts the interference coefficient and the incremental adjustment factor based on the current parameter state and the change rate of the loss function.

[0116] In some embodiments, the adjustment of the force field distribution includes:

[0117] Calculate the initial force field distribution function, combine the nonlinear mapping of weights and biases, and use an activation function to achieve the nonlinear mapping of weights and biases;

[0118] Dynamically adjust the interference coefficient according to the change rate of the loss function to suppress parameter oscillation;

[0119] Adaptive adjustment of the learning rate through the incremental adjustment factor and the perturbation correction factor.

[0120] In some embodiments, the output categories of the adaptability evaluation model include low comfort, medium comfort, and high comfort.

[0121] In some embodiments, the training termination condition of the adaptive force field optimization mechanism is to reach the preset maximum number of iterations or the change rate of the loss function is lower than the threshold.

[0122] Specifically, the trained adaptability evaluation model is a neural network based on adaptive force field optimization. To solve the technical problem that the neural network training process is prone to falling into local optimal solutions, the neural network based on adaptive force field optimization simulates the adaptive force field phenomenon in nature, adopts an optimization mechanism similar to force field interference, and combines an adaptive adjustment method to replace the traditional gradient descent method, so as to find the global optimal solution in the search space and be able to adaptively adjust the update strength of each parameter, avoiding the problem of local optimal solutions and accelerating the convergence of the neural network.

[0123] Specifically, the training process of the neural network based on adaptive force field optimization is as follows:

[0124] 1. Initialize the weights and biases of the neural network, and set the distribution function of the initial force field to construct the initial state of the algorithm, which is expressed as:

[0125]

[0126] In the formula, is the initial weight of the neural network, is the initial bias of the neural network, is a normal distribution with a mean of 0 and a variance of σ pa 2 and σ pa is the standard deviation for initializing the neural network parameters.

[0127] Preferably, σ pa is set to 0.01.

[0128] Furthermore, construct the initial force field distribution function. The force field distribution function simulates the action of the force field and dynamically adjusts the distribution of its parameters according to the current state of the neural network parameters at each iteration. Through the dynamic adjustment mechanism of the force field, the parameter update strength of different neurons can be differentially processed according to the feature complexity of the building design data, and complex and subtle building design features can be extracted more precisely. The calculation method is:

[0129]

[0130] In the formula, is the initial force field distribution function, α0 is the initial interference coefficient, p is a constant to prevent division by zero; is a non-linear mapping function related to the initial weights and biases.

[0131] Preferably, the specific implementation method is,

[0132] Preferably, p is set to 10 -6 .

[0133] 2. Calculate the force field according to the weights and biases of the current neural network, and make the weight update strength of different nodes different through the dynamic adjustment of the force field, so as to solve the technical problem of differential processing of different parameters in the training of the neural network, and more precisely control the update direction and amplitude of each parameter of the neural network, which is expressed as:

[0134]

[0135] In the formula, is the force field function at the t-th iteration, is the value of the i-th weight of the neural network at the t-th iteration, is the value of the i-th bias of the neural network at the t-th iteration, N pm is the number of neurons in the neural network, α i is the interference coefficient of the i-th neuron of the neural network; is the non-linear mapping function for the i-th neuron.

[0136] Preferably, the specific implementation is as follows,

[0137] Furthermore, to prevent the weights or biases from being too large and causing unstable training, the interference coefficient of the neurons is exponentially decayed. By suppressing unnecessary parameter oscillations, the stability of the neural network model is enhanced, making the results of complex building design feature extraction more reliable. The calculation method is as follows:

[0138]

[0139] In the formula, α i is the dynamic adjustment coefficient of the i-th neuron of the neural network, θ gas is the adjustment parameter of the interference coefficient, σ i is the adjustment standard deviation of the interference coefficient.

[0140] Preferably, θ gas is set to 0.2, σ gas is set to 0.1.

[0141] 3. Calculate the update amount of each parameter through the force field function, solve the problem that single linear superposition cannot fully utilize the multi-parameter interaction information, and perform non-linear correction on the parameters of the neural network, so as to handle the multi-dimensional feature interaction problem in building design data, which is expressed as:

[0142]

[0143] In the formula, is the update amount of the weight at the t-th iteration of the neural network, λ p is the adjusted learning rate of the neural network, Npe is the number of samples input to the neural network in the current batch, is the force field function of the neural network input for the i-th sample at the t-th iteration, is the incremental adjustment factor at the t-th iteration of the neural network.

[0144] Preferably, λ p is set to 0.01.

[0145] Furthermore, the incremental adjustment factor adaptively adjusts the learning rate according to the network state and loss changes, thus avoiding the defect of over-reliance on a fixed learning rate in traditional methods. By dynamically adjusting the learning rate through the perturbation correction factor and the incremental adjustment factor, it can adjust the optimization intensity according to the change rate of data features, thereby enhancing the adaptability of the model to different building design data features. The calculation method is as follows:

[0146]

[0147] In the formula, is the perturbation correction factor of the neural network at the t-th iteration; α utr is the incremental adjustment factor adjustment constant.

[0148] Preferably, set α utr = 0.05; is the change amount of the loss function of the neural network at the t-th iteration, is the loss function value of the neural network at the (t - 1)-th iteration.

[0149] Furthermore, the perturbation correction factor non-linearly adjusts the magnitude of the gradient by combining the change of the loss function with the current weights. The magnitude of the perturbation factor is inversely proportional to the square of the change rate of the loss function and directly proportional to the magnitude of all current weights (i.e., the complexity of the network). It adjusts the gradient according to the oscillation degree of the current network state, enabling the network to adapt to different optimization stages. The calculation method is as follows:

[0150]

[0151] In the formula, α rds is the adjustment parameter of the perturbation factor.

[0152] Preferably, set α rds = 0.2, which is used to control the intensity of perturbation correction; is the sum of the absolute values of all current weights, representing the overall complexity of the network.

[0153] When is relatively large, it indicates that the change of the loss function is relatively drastic. At this time, the perturbation correction factor will become smaller, so that the magnitude of gradient update is suppressed, avoiding oscillation or overfitting of the model due to too fast weight adjustment during the neural network training process;

[0154] When is relatively small, indicating that the change of the loss function is relatively stable, the perturbation correction factor will approach 1, enabling the weight update to maintain a relatively large magnitude, thereby accelerating the convergence of the network.

[0155] 4. Through an adaptive feedback mechanism, the distribution of the force field is dynamically changed according to the change rate of the loss function, thereby solving the problem that the neural network cannot adaptively increase or decrease the interference intensity during the training process, which is expressed as:

[0156]

[0157] In the formula, is the change amount of the loss function of the neural network at the t-th iteration, is the loss function value of the neural network at the t-th iteration, is the loss function value of the neural network at the (t - 1)-th iteration.

[0158] Furthermore, in order to adaptively adjust the force field, the force field is corrected according to the change amount of the loss function, and the calculation method is expressed as:

[0159]

[0160] In the formula, is the force field distribution at the (t + 1)-th iteration, is the force field function at the t-th iteration; δ p is the feedback adjustment coefficient.

[0161] Preferably, δ p is set to 0.1; p is a constant to prevent division by zero.

[0162] Preferably, p is set to 10 -6 .

[0163] 5. According to the non-linear correction result, the weights of the neural network are updated, continuously optimized and gradually approximated to the global optimal solution, and the calculation method is expressed as:

[0164]

[0165] In the formula, is the weight of the neural network after the (t + 1)-th iteration.

[0166] 6. Repeat the above steps iteratively until the preset iteration stop condition is satisfied, which means the model training is completed.

[0167] In a specific embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations.

[0168] Preferably, the preset maximum number of iterations is set to 1000 times.

[0169] After the neural network training is completed, the building design data features extracted by the neural network are classified using a preset classification tool. The preset classification tool can be any one of the Softmax function, random forest classification area, decision tree classification area, and support vector machine classifier, and the classification categories include:

[0170] 0 - Low comfort level (not suitable for long-term residence or use);

[0171] 1 - Medium comfort level (basically meeting the living or use needs);

[0172] 2 - High comfort level (optimizing the living or use experience).

[0173] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. An adaptive evaluation method for virtual reality building design, characterized in that Including: Collecting architectural design data through a virtual reality environment, where the architectural design data includes user interaction behavior paths, geometric form parameters, material property parameters, and environmental simulation parameters; Dynamically updating the adaptability evaluation model using an incremental learning framework, including knowledge retention based on historical tasks and incremental learning of new tasks; Augmenting the original architectural design data using a generative adversarial network and enhancing the diversity and quality of the generated data through a composite loss function; Training a neural network based on an adaptive force field optimization mechanism to dynamically adjust the intensity of parameter updates; Conducting an adaptability evaluation of the trained model and outputting the comfort level of the design scheme under different scenarios.

2. The method according to claim 1, characterized in that, The collection method of the architectural design data includes: Automated export and semi-automated interaction recording; The automated export extracts data from the design software automatically through a design software tool; the semi-automated interaction recording realizes the recording and collation of data through the combination of user interaction and data processing algorithms.

3. The method according to claim 1, characterized in that The specific steps of the incremental learning framework include: Judging whether the accuracy rate of the new dataset on the validation set of the original data has increased by more than a preset threshold. If satisfied, the old and new data are weighted and fused by category; Labeling the unlabeled new data through the pseudo-label method to generate an extended training set; Iteratively updating the model parameters until convergence.

4. The method according to claim 1, wherein The input of the conditional generator of the generative adversarial network includes random noise and the feature statistical information of the architectural design data.

5. The method according to claim 1, wherein The composite loss function includes adversarial loss, content loss, and feature loss, where the content loss constrains the pixel space consistency between the generated data and the real data, and the feature loss constrains the consistency in the high-level feature space.

6. The method according to claim 1, characterized in that, The adaptive force field optimization mechanism adjusts the update amplitude of each parameter through dynamic force field distribution, and the force field distribution adaptively adjusts the interference coefficient and the incremental adjustment factor based on the current parameter state and the change rate of the loss function.

7. The method according to claim 6, characterized in that, The adjustment of the force field distribution includes: Calculating the initial force field distribution function, combining the nonlinear mapping of weights and biases, and using an activation function to realize the nonlinear mapping of weights and biases; Dynamically adjusting the interference coefficient according to the change rate of the loss function to suppress parameter oscillation; Adapting the learning rate through the incremental adjustment factor and the perturbation correction factor.

8. The method according to claim 1, characterized in that The output categories of the adaptability evaluation model include low comfort, medium comfort, and high comfort.

9. The method according to claim 1, wherein The generative adversarial network gradually increases the weights of the content loss and the feature loss during the training process. In the initial stage, the adversarial loss is the main one, and the regulatory roles of the content and feature losses are gradually enhanced in the later stage.

10. The method according to claim 1, wherein The training termination condition of the adaptive force field optimization mechanism is to reach the preset maximum number of iterations or the change rate of the loss function is lower than the threshold.

Images